最新刊期
卷 34 ， 期 12 ， 2026
-
High-precision optical measurement system for dynamic angle based on laser differential alignment AI导读
摘要：To address the degradation of alignment measurement accuracy caused by the difficulty in suppressing optical aberrations in dynamic angle measurement systems， this paper proposed a high-precision optical measurement technique featuring dynamic aberration suppression based on the principle of laser differential alignment. A differential alignment mathematical model incorporating aberration factors was established to thoroughly analyze the impact of aberrations on the energy distribution of the detection spot and the zero-point positioning of the differential signal. This analysis facilitated the derivation of a quantitative relationship between wavefront aberration and angular measurement sensitivity. Subsequently， a long-focal-length catadioptric optical structure alongside even aspheric surfaces was employed to effectively suppress both on-axis and off-axis residual aberrations， thereby enhancing the angular measurement accuracy. Theoretical analysis and preliminary experiments demonstrate that the proposed technique can effectively suppress optical aberrations. Across the full field of view （FOV）， the root-mean-square （RMS） spot radii are all smaller than the Airy disk radius， the wavefront RMS error is less than 0.0424λ， and the modulation transfer function （MTF） approaches the diffraction limit. The alignment measurement repeatability of the proposed system reaches 0.027 arcseconds， which is 41% higher than that of the widely adopted ELCOMAT®3000 autocollimator， and the maximum alignment measurement deviation of this system is only 0.085 arcseconds. This study provides a reliable technical pathway for high-precision dynamic angle measurement.关键词：optical design;differential alignment;high precision;aberration suppression73|9|0更新时间：2026-07-20 - 摘要：Raman spectroscopy enables multi-component gas analysis but faces challenges in complex industrial environments such as converter and coke oven gases. In these settings， online measurements of H2， CO， and CH4 suffer from reduced accuracy due to multi-source interferences， including gas pressure fluctuations， laser power drift， and variations in optical path transmission efficiency. To address these issues， this paper proposed a Raman spectroscopic analysis method based on internal standard calibration and multivariate joint modeling. First， CO2， which was ubiquitous and chemically stable in the gas mixture， was used as an internal reference component to construct the internal standard ratio Ri， thereby eliminating common-mode interference in system gain. Second， a system gain calibration operator Sfactor was defined to enable real-time characterization of system state variables， including pressure and scattered light intensity. Finally， several parameters were jointly used as input features， including the internal standard ratio， gas pressure， scattered light intensity， and the system gain calibration operator. A multivariate nonlinear regression model was then established using the LightGBM algorithm. This model enabled high-precision retrieval of H2， CO， and CH4 concentrations. Experimental results demonstrate that the proposed multivariate joint model achieves a root mean square error （RMSE） of below 0.09% for each of the three components. The mean absolute percentage error （MAPE） for each component is below 0.6%. Compared to models using only raw spectral data or a single internal standard ratio， the RMSE of the proposed model is reduced by approximately 37% to 63%. This method effectively compensates for the nonlinear interferences caused by variations in gas pressure and scattered light intensity， significantly improving the accuracy of concentration retrieval for multi-component gases in complex industrial conditions using Raman spectroscopy. and provides a feasible solution for online industrial gas detection.关键词：Raman spectroscopy;coal gas;internal standard;multivariate joint modeling;machine learning32|7|0更新时间：2026-07-20
- 摘要：To meet the testing requirements for the overall target surface response non-uniformity and inter-pixel signal crosstalk of infrared focal plane detectors， this paper designed an infrared focal plane spatial performance parameter adaptive detection system. The system employed a 1 550 nm laser as the light source， used a Galilean aspheric lens group to shape the Gaussian beam into a flat-top beam， and then expanded the flat-top beam through a Keplerian beam expander to obtain a homogenized beam with a diameter greater than 25 mm and uniformity exceeding 96%， fulfilling the testing needs for the target surface response non-uniformity of infrared focal plane detectors. To generate high-quality small light spots， the system utilized an adaptive optical system to correct wavefront aberrations， and after focusing through the objective lens， a small light spot with a diameter less than 5 μm was obtained， meeting the precise measurement requirements for inter-pixel signal crosstalk in infrared focal plane detectors. Experimental results indicate that the uniformity of homogenized beam reaches 96.08%， and the diameter of the light spot is 4.4 μm. This system integrates both functions into one unit， providing an effective experimental tool for performance testing of infrared focal plane detectors.关键词：homogenized beam;infrared small spot;aspheric lens group;adaptive optics;infrared focal plane detector48|8|0更新时间：2026-07-20
Modern Applied Optics
- 摘要：Measuring the dimensions and assembly pose of tiny circular coils is crucial for ensuring the performance of precision electromechanical products. However， pseudo-edge interference and field-of-view limitations significantly impact the measurement process. This paper proposed a cross-scale precision measurement method that integrated semantic segmentation. First， a U-Net++ network integrating Efficient Channel Attention （ECA） and edge enhancement branches was constructed to achieve robust segmentation and coarse localization of the coil region. Then， within the region of interest constrained by the segmentation mask， sub-pixel edge points were extracted using 8th-order Zernike moments. High-precision reconstruction of the coil contour was achieved through four-field-of-view coordinate unification and Levenberg-Marquardt （LM） circle fitting. Finally， the established measurement method was integrated into an automated coil assembly equipment for automated assembly application experiments. Experimental results show that the improved model achieves Dice coefficients， mIoU， and Precision of 94.95%， 90.46%， and 95.27%， respectively， in the coil microscopic image segmentation task， demonstrating good anti-interference capabilities. Furthermore， in 10 assembly experiments， the average coaxiality errors of the upper and lower coils were 12.6 μm and 13.3 μm， respectively， which is about 43% lower than other commonly used methods. The number of assembly samples that met the standards increased from 3 to 9. This demonstrates that the method proposed in this paper has good robustness and measurement accuracy， providing a new approach for visual guidance in the automated assembly of micro-coil components.关键词：micro-assembly;microscopic vision measurement;semantic segmentation;micro circular coil36|7|0更新时间：2026-07-20
- 摘要：Aimed at the high-precision milling and grinding problems of complex special-shaped optical prisms， a special high-precision milling and grinding machine tool based on the HNC-808DiG numerical control system was studied. The machine tool realized continuous milling and grinding of multiple surfaces through the coordinated control of the cup wheel rotating around the Z-axis + controlled linear motion along the Z-axis and the workpiece rotating around the B-axis + controlled linear motion along the X-axis. A non-concentric arc corner trajectory generation algorithm with a three-axis linkage controlled by a custom macro program was adopted to achieve the milling and grinding of smooth transition surfaces between adjacent surfaces. The effectiveness and accuracy of the theoretical research were verified through milling and grinding experiments on two typical complex special-shaped optical prisms： a pentaprism and a triangular prism. The experimental results show that： For the pentaprism， the dimensional error ≤ 0.02 mm， angular error ≤ 20.5″， and edge thickness difference ≤ 0.008 mm； For the triangular prism， the dimensional error ≤ 0.03 mm， angular error ≤ 26.5″， and edge thickness difference ≤ 0.012 mm. The transition arc corners exhibit good radius consistency and smooth continuous surfaces， meeting assembly requirements. The results demonstrate that the special prism milling and grinding machine developed in this paper based on the HNC-808DiG standard CNC system can effectively solve the high-precision machining problems of complex special-shaped optical prisms， and provides a feasible engineering solution for the high-precision milling and grinding of domestic complex special-shaped optical prisms.关键词：complex special-shaped optical prism;cup wheel;three-axis linked milling and grinding;non-concentric arc corner;trajectory modeling and verification19|6|0更新时间：2026-07-20
- 摘要：Thermal error is a critical factor affecting the motion accuracy of the direct-drive feed axis. Conventional mechanistic models often exhibit low prediction accuracy due to the complexity of modeling and solving； meanwhile， empirical models rely excessively on high-quality data and lack physical interpretability. This paper proposes a thermal error modeling method based on a Gated Mixture Physics-Informed Neural Network （GM-PINN）. First， an analysis of the thermal behavior during the operation of the direct-drive feed axis is conducted to derive and establish the governing temperature control partial differential equations， providing physical constraints for the model. Second， the GM-PINN model is constructed using a "Network Division—Gated Fusion" strategy. This strategy adaptively adjusts the integration of data-driven and physics-driven components according to data features， effectively mitigating the training competition between the two. Experimental results demonstrate that this method successfully integrates the advantages of mechanistic analysis and data-driven approaches. Under low， medium， and high-speed operating conditions， the prediction errors remain within ［-0.42， 0.04］ μm， ［-0.03， 0.74］ μm， and ［-0.79， 1.85］ μm， respectively. The proposed method possesses strong physical interpretability and is particularly suitable for thermal error modeling where observational data are limited but the underlying physical mechanisms are well-defined.关键词：GM-PINN;direct-drive feed axis;thermal errors;physical interpretability24|8|0更新时间：2026-07-20
- 摘要：For precision assembly tasks with high requirements for coaxial alignment accuracy， a dual-field telecentric vision measurement system was built. Different from traditional multi-camera or stereo vision systems， the proposed system has no overlapping field of view， making it difficult to directly establish the relationship between the two telecentric imaging units. To solve this problem， a coaxial alignment method for a dual-field telecentric vision system was proposed. A calibration plate was moved into the fields of view of the two telecentric imaging units in sequence to establish the relationship between them， and a mathematical model for coaxial alignment was built. On this basis， taking the armature assembly as the research object， the main error sources affecting coaxial alignment accuracy were analyzed. Error modeling was carried out based on the Abbe-Bryan principle， and corresponding error compensation methods were proposed. Indentation experiment results show that the coaxial alignment deviations were controlled within 5 μm. In addition， press-fitting experiments of the armature assembly further verified the feasibility and effectiveness of the proposed method for precision assembly of shaft-hole components.关键词：telecentric imaging;precision assembly;error compensation;coaxial alignment;dual-field32|7|0更新时间：2026-07-20
Micro\/Nano Technology and Fine Mechanics
- 摘要：To address the issue that existing tracking methods could not achieve decoupling and synergy between temporal features and spatial features， a Dual-Guided Feature Enhancement Mamba for Visual Object Tracking was proposed. Built upon the Vision Mamba backbone， a Spatial-Temporal Decoupling and Collaboration Module was designed to separate and bidirectionally enhance coupled spatiotemporal information. A dual-branch imitation attention mechanism was constructed， enabling shallow features to emulate the focus of deep features， thereby enhancing discriminative capability. A dynamic template enhancement and update strategy was introduced， leveraging collaborative long- and short-term memory banks to achieve adaptive template iteration. Experimental results demonstrate that the proposed algorithm achieves a precision of 93.7% and a success rate of 75.6% on the OTB100 dataset. It also attains leading or suboptimal performance on most metrics across the LaSOT， TrackingNet， and GOT-10k benchmarks， with an AUC of 75.8% on LaSOT， 86.2% on TrackingNet， and an average overlap of 78.9% on GOT-10k. The algorithm effectively balances and synergizes temporal features with spatial characteristics， outperforming mainstream comparison methods and demonstrating strong accuracy and robustness under complex challenges.关键词：object tracking;Mamba;spatiotemporal coupling;imitation attention;dynamic template25|7|0更新时间：2026-07-20
- 摘要：Data-driven deep learning methods that decouple the geometry and texture of 3D objects enable appearance editing. However， the view fidelity rendered by neural networks remains insufficient after local texture editing， and controllability is still weak due to implicit end-to-end view rendering. To flexibly edit object appearance as in traditional methods， this paper proposed an editable appearance neural texture mapping method that decoupled diffuse and specular reflections. First， a three-stage independent learning framework was constructed based on the physical illumination model， including diffuse reflection， specular reflection， and texture mapping. Next， a neural radiance field was built via spherical spatial sampling， and a loss function for the diffuse reflection component was established to learn the diffuse illumination energy distribution while conducting supervised learning on real-world data. Then， the neural shell network method based on 3D Gaussian splatting employed multi-resolution hash encoding for learnable color information； after incorporating the diffuse reflection component， supervised learning with real images was performed to achieve view-dependent specular reflection modeling. Finally， a cycle-consistent texture mapping network was used to establish position mappings independent of the color space to obtain explicit 2D textures， and the view-independent diffuse color components were mapped one-to-one to texture atlases， realizing the inverse operation of editing texture images for neural textures. Experimental results show that on the public DTU dataset， compared with the state-of-the-art editable texture methods NeuTex and Texture-GS， the peak signal-to-noise ratio （PSNR） on the novel view synthesis test set is improved by an average of 5.6%， and the perceptual image loss is reduced by an average of 18%. It also achieves superior structural similarity on the NeRFSyn dataset. Moreover， the multi-view images rendered by the neural network after explicit texture editing are significantly clearer. The proposed method realizes the construction and controllable editing of neural textures， enabling flexible modification of the appearance of 3D models with neural implicit representations directly on 2D textures while ensuring multi-view consistency of appearance.关键词：physically based rendering;neural radiance field;3D Gaussian splatting;texture mapping17|7|0更新时间：2026-07-20
- 摘要：High-precision calibration of large field-of-view cameras remains challenging in machine vision， owing to the difficulties in fabricating， storing， transporting and handling large high-precision calibration targets. To address this issue， this paper presented a high-precision calibration method for large field-of-view cameras using reinforcement learning. Based on the Asynchronous Advantage Actor-Critic （A3C） algorithm， the proposed method optimized small-target poses for Zhang's calibration under a large field of view， and obtained the optimal sample distribution of equivalent large-target images and features， thus improving calibration accuracy. The target pose optimization problem was formulated as a Markov game， and N parallel target agents were trained via the A3C algorithm in separate environment instances to optimize target poses. The learned policy was then applied to large field-of-view camera calibration. Experiments on a 10-square-meter field of view show that the reprojection error is stably below 0.2 mm， Compared with the fixed defocus method， accuracy improved by about 50%， and compared with the MAPPO algorithm， accuracy improved by about 12.3%. The proposed method has been successfully applied to full-scene motion parameter measurement of rotor blades in rotor wind tunnel tests， meeting the high-precision and high-stability requirements of camera calibration.关键词：large field of view;target pose optimization;multi-agent reinforcement learning;curve fitting;Markov game21|7|0更新时间：2026-07-20
- 摘要：In order to address the demand for open-ended text-instruction-guided fusion in infrared–visible image fusion tasks， this paper proposed an infrared–visible image fusion framework based on learnable frequency-band decomposition and text-guided gating. Specifically， the proposed method constructed a text–image alignment module by jointly exploiting Contrastive Language–Image Pre-training （CLIP） and the Grounded Segment Anything Model （Grounded SAM）， which associated textual semantic priors with local image regions. Then， a dedicated learnable frequency-band decomposition module was introduced for bimodal feature extraction， which adaptively separated the low-frequency base representing global structures from the high-frequency details describing fine textures. Based on this decomposition， a text-aware gated fusion module spatially regulated the feature contribution weights of the infrared and visible modalities. In addition， a multi-constraint loss function was formulated to incorporate gradient fidelity， and gating-map regularization， thereby preserving structural integrity and textual semantic consistency in the fused results. Experimental results on the Low-Light Visible-Infrared Paired （LLVIP） dataset demonstrate that， compared with TextFusion， the proposed method improves the text-related metrics edge preservation index plus （Qabf+） and Structural Similarity Index Measure plus （SSIM+） by 0.070 5 and 0.036 4， respectively. For common image fusion metrics， Mutual Information （MI） and Peak Signal-to-Noise Ratio （PSNR） are improved by 1.516 4 and 1.756 6 dB， respectively. Moreover， the proposed method achieves a 0.9% improvement in mean Average Precision over Intersection over Union thresholds from 0.5 to 0.95 （mAP@0.5∶0.95） in the downstream object detection task. These results indicate that textual instructions can effectively guide the image fusion process in practical scenarios， thereby achieving semantically controllable infrared–visible image fusion.关键词：image fusion;text guidance;infrared image;visible image;frequency-band decomposition24|7|0更新时间：2026-07-20
Information Sciences
- Address：No.3888 Dong Nanhu Road, Changchun, Jilin, China Postal code：130033
- Tel：0431-86176855 Email：gxjmgc@ciomp.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 吉ICP备11002662号-17
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0